Problem: An application violating the architectural rules regarding
operation dependencies and having specific Register Stack Engine (RSE)
state at the time of the violation, may result in an illegal operation
fault and invalid RSE state. Such faults may initiate a cascade of
repeated illegal operation faults within OS interruption handlers.
The specific behavior is OS dependent.
Implication: An application causing an illegal operation fault with
specific RSE state may result in a series of illegal operation faults
and an eventual OS stack overflow condition.
Workaround: OS interruption handlers that switch to kernel backing
store implement a check for invalid RSE state to avoid the series
of illegal operation faults.
The core of the workaround is the RSE_WORKAROUND code sequence
inserted into each invocation of the SAVE_MIN_WITH_COVER and
SAVE_MIN_WITH_COVER_R19 macros. This sequence includes hard-coded
constants that depend on the number of stacked physical registers
being 96. The rest of this patch consists of code to disable this
workaround should this not be the case (with the presumption that
if a future Itanium processor increases the number of registers, it
would also remove the need for this patch).
Move the start of the RBS up to a mod32 boundary to avoid some
corner cases.
The dispatch_illegal_op_fault code outgrew the spot it was
squatting in when built with this patch and CONFIG_VIRT_CPU_ACCOUNTING=y
Move it out to the end of the ivt.
Signed-off-by: Tony Luck <tony.luck@intel.com>
This patch implements VIRT_CPU_ACCOUNTING for ia64,
which enable us to use more accurate cpu time accounting.
The VIRT_CPU_ACCOUNTING is an item of kernel config, which s390
and powerpc arch have. By turning this config on, these archs
change the mechanism of cpu time accounting from tick-sampling
based one to state-transition based one.
The state-transition based accounting is done by checking time
(cycle counter in processor) at every state-transition point,
such as entrance/exit of kernel, interrupt, softirq etc.
The difference between point to point is the actual time consumed
during in the state. There is no doubt about that this value is
more accurate than that of tick-sampling based accounting.
Signed-off-by: Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Instead of pinning per-cpu TLB into a DTR, use DTC. This will free up
one TLB entry for application, or even kernel if access pattern to
per-cpu data area has high temporal locality.
Since per-cpu is mapped at the top of region 7 address, we just need to
add special case in alt_dtlb_miss. The physical address of per-cpu data
is already conveniently stored in IA64_KR(PER_CPU_DATA). Latency for
alt_dtlb_miss is not affected as we can hide all the latency. It was
measured that alt_dtlb_miss handler has 23 cycles latency before and
after the patch.
The performance effect is massive for applications that put lots of tlb
pressure on CPU. Workload environment like database online transaction
processing or application uses tera-byte of memory would benefit the most.
Measurement with industry standard database benchmark shown an upward
of 1.6% gain. While smaller workloads like cpu, java also showing small
improvement.
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Memory errors encountered by user applications may surface
when the CPU is running in kernel context. The current code
will not attempt recovery if the MCA surfaces in kernel
context (privilage mode 0). This patch adds a check for cases
where the user initiated the load that surfaces in kernel
interrupt code.
An example is a user process lauching a load from memory
and the data in memory had bad ECC. Before the bad data
gets to the CPU register, and interrupt comes in. The
code jumps to the IVT interrupt entry point and begins
execution in kernel context. The process of saving the
user registers (SAVE_REST) causes the bad data to be loaded
into a CPU register, triggering the MCA. The MCA surfaces in
kernel context, even though the load was initiated from
user context.
As suggested by David and Tony, this patch uses an exception
table like approach, puting the tagged recovery addresses in
a searchable table. One difference from the exception table
is that MCAs do not surface in precise places (such as with
a TLB miss), so instead of tagging specific instructions,
address ranges are registers. A single macro is used to do
the tagging, with the input parameter being the label
of the starting address and the macro being the ending
address. This limits clutter in the code.
This patch only tags one spot, the interrupt ivt entry.
Testing showed that spot to be a "heavy hitter" with
MCAs surfacing while saving user registers. Other spots
can be added as needed by adding a single macro.
Signed-off-by: Russ Anderson (rja@sgi.com)
Signed-off-by: Tony Luck <tony.luck@intel.com>
A pte may be zapped by the swapper, exiting process, unmapping or page
migration while the accessed or dirty bit handers are about to run. In that
case the accessed bit or dirty is set on an zeroed pte which leads the VM to
conclude that this is a swap pte. This may lead to
- Messages from the vm like
swap_free: Bad swap file entry 4000000000000000
- Processes being aborted
swap_dup: Bad swap file entry 4000000000000000
VM: killing process ....
Page migration is particular suitable for the creation of this race since
it needs to remove and restore page table entries.
The fix here is to check for the present bit and simply not update
the pte if the page is not present anymore. If the page is not present
then the fault handler should run next which will take care of the problem
by bringing the page back and then mark the page dirty or move it onto the
active list.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
unaligned_access does fetch cr.ipsr, then calls
dispatch_unaligned_handler, but dispatch_unaligned_handler fetches
cr.ipsr again, so delete the first one.
Signed-off-by: Zhang Yanmin <yanmin.zhang@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Polish the comments specifically in vhpt_miss and nested_dtlb_miss
handlers. I think it's better to explicitly name each page table
level with its name instead of numerically name them. i.e., use
pgd, pud, pmd, and pte instead of referring as L1, L2, L3 etc.
Along the line, remove some magic number in the comments like:
"PTA + (((IFA(61,63) << 7) | IFA(33,39))*8)". No code change at
all, pure comment update. Feel free to shoot anything you have,
darts or tomahawk cruise missile. I will duck behind a bunker ;-)
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Acked-by: Robin Holt <holt@sgi.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
From source code inspection, I think there is a bug with 4 level
page table with vhpt_miss handler. In the code path of rechecking
page table entry against previously read value after tlb insertion,
*pte value in register r18 was overwritten with value newly read
from pud pointer, render the check of new *pte against previous
*pte completely wrong. Though the bug is none fatal and the penalty
is to purge the entry and retry. For functional correctness, it
should be fixed. The fix is to use a different register so new
*pud don't trash *pte. (btw, the comments in the cmp statement is
wrong as well, which I will address in the next patch).
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
This patch introduces 4-level page tables to ia64. I have run
some benchmarks and found nothing interesting. Performance has
consistently fallen within the noise range.
It also introduces a config option (setting the default to 3
levels). The config option prevents having 4 level page
tables with 64k base page size.
Signed-off-by: Robin Holt <holt@sgi.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Delete obsolete stuff from arch Makefile
Rename file to asm-offsets.h
The trick used in the arch Makefile to circumvent the circular
dependency is kept.
Signed-off-by: Sam Ravnborg <sam@ravnborg.org>
This patch greatly speeds up the handling of lfetch.fault instructions
which result in NaT consumption. Due to the NaT-page mapped at address
0, this is guaranteed to happen when lfetch.fault'ing a NULL pointer.
With this patch in place, we can even define prefetch()/prefetchw() as
lfetch.fault without significant performance degradation. More
importantly, it allows compilers to be more aggressive with using
lfetch.fault on pointers that might be NULL.
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
The nested_dtlb_miss handler currently does not handle fault from
hugetlb address correctly. It walks the page table assuming PAGE_SIZE.
Thus when taking a fault triggered from hugetlb address, it would not
calculate the pgd/pmd/pte address correctly and thus result an incorrect
invocation of ia64_do_page_fault(). In there, kernel will signal SIGBUS
and application dies (The faulting address is perfectly legal and we
have a valid pte for the corresponding user hugetlb address as well).
This patch fix the described kernel bug. Since nested_dtlb_miss is a
rare event and a slow path anyway, I'm making the change without #ifdef
CONFIG_HUGETLB_PAGE for code readability. Tony, please apply.
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
This patch reorganizes break_fault() to optimistically assume that a
system-call is being performed from user-space (which is almost always
the case). If it turns out that (a) we're not being called due to a
system call or (b) we're being called from within the kernel, we fixup
the no-longer-valid assumptions in non_syscall() and .break_fixup(),
respectively.
With this approach, there are 3 major phases:
- Phase 1: Read various control & application registers, in
particular the current task pointer from AR.K6.
- Phase 2: Do all memory loads (load system-call entry,
load current_thread_info()->flags, prefetch
kernel register-backing store) and switch
to kernel register-stack.
- Phase 3: Call ia64_syscall_setup() and invoke
syscall-handler.
Good for 26-30 cycles of improvement on break-based syscall-path.
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Using stf8 seemed like a clever idea at the time, but stf8 forces
the cache-line to be invalidated in the L1D (if it happens to be
there already). This patch eliminates a guaranteed L1D cache-miss
and, by itself, is good for a 1-2 cycle improvement for heavy-weight
syscalls.
Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
Tony Luck
Hidetoshi Seto
Chen, Kenneth W
Jörn Engel
Russ Anderson
Christoph Lameter
Zhang, Yanmin
Robin Holt
Keith Owens
Sam Ravnborg
David Mosberger-Tang
Linus Torvalds